Biased but Not Unfair vs Fair but Biased
In the article, under section Q: Why might a pricing model be biased but not unfair (or vice versa)?, it implies that:
1. Biased but Not unfair -> Possible
And then says that:
2. Fair but Biased -> Not typically possible
This seems rather contradictory to me. Isn't "fair" basically the same as "not unfair"?
The justification provided for the second statement is also the same as the key insight explaining that a model can be biased without being unfair:
Comments
In the context of the paper -> fair at a very high level means that insureds are not disproportionately affected or harmed by the pricing process. Thus if a model is fair, any differences in rates should not harm the policyholder since those differences would be justified, nor would they be disproportionately be affected.
An example of biased but fair would be for example, when we build a pricing GLM. If we have data indicating that older drivers (75+) have better loss experience than drivers aged 40, we will probably disregard the data and manually fit a curve as we know from experience that this is not true. Your model is now biased, but not necessarily unfair
What I meant is that the wiki article seems to differentiate between 'Biased and Not unfair' and 'Biased and Fair', but isn't 'not unfair' the same as 'fair'?
I have spent some time thinking about it and came up with new examples for each of the bias-fairness combinations. I also added a key insight that is a reinforcement of the key insight from PQ1. Thanks for pointing that out